The importance of developing data science workflows enhances productivity by optimizing cost and improving user experience. The data going through an iterative development makes the workflow achieve data exploration and model prototyping.

The data science workflow comprises 3 different skill sets and personas,

1.  Data Engineer - Responsible for Data ingestion, storage and cleansing of the data to make data ready for the data scientists
2. Data Scientist – The Data scientist gets the cleaned, authentic and quality data to develop features, build model prototypes and test the model on real datasets on various algorithms to increase accuracy depending on the business use case. The result of this step will be the model ready for the production environment
3. ML Engineers – Their role is to operationalize data processing into production, deploying models created out of data into the production system. In production ML engineers constantly monitor production model performance and accuracy.

It is found that to improve productivity, around 90% of the time is spent in experimentation, data exploration, and model prototyping stages. Data science workflow is iterated to achieve feature engineering, model selection, and hyper-parameters to finally select a model that meets all requirements and goals for the business use cases.

Limitations with traditional development setup with CPU workstation or on Public Cloud 

1. Higher cloud operational costs on data training and experimentation
2. Lack of resource availability or waiting time on CPU workstation and limited compute processing
3. Security and vendor lock-in on training data in the cloud environment
4.  Infra support and maintenance on the cloud incur operational costs

So, to improve productivity and operational costs, the data exploration and model prototyping can happen on a GPU-accelerated setup on a local workstation and the other two steps, full model training and model scoring with real data, can take place remotely as per business requirement.

Advantages of GPU-accelerated local workstation for data science workflows

1. Numerous experimentations on model prototyping
2. GPU-accelerated workstation has 20x more power than CPU for processing large datasets in training
3. Reducing cloud cost and a positive ROI on a local workstation 
4. Increased productivity in processing complex data for model building and the model can be tested on real data on production remotely

The Data Science Stack consists of Drivers, CUDA-X, and GPU-accelerated SDKs and frameworks. The NGC (Nvidia GPU Cloud) simplifies and accelerates end-to-end workflows, which are comprised of,

1. Containers for high-performance computing, deep learning, and Machine learning
2. Pre-trained models for NLP, Vision, DLRM and many others
3. Industry application frameworks like Clara, Jarvis, Issac
4. Helm Charts for Triton inference server, GPU operator on Kubernetes Cluster
5. It serves the purpose of hosting on On-premises, Cloud, Hybrid-cloud or edge infrastructure